expert systems with applications - ایران...

Expert Systems with Applications 40 (2013) 4241–4251

Contents lists available at SciVerse ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier .com/locate /eswa

More than words: Social networks’ text mining for consumer brand sentiments

Mohamed M. MostafaInstituto Universitário de Lisboa, Business Research Unit, Avenida das Forças Armadas, Lisbon, Portugal

a r t i c l e i n f o a b s t r a c t

Keywords:Consumer behaviorGlobal brandsSentiment analysisText miningTwitter

0957-4174/$ - see front matter � 2013 Elsevier Ltd. Ahttp://dx.doi.org/10.1016/j.eswa.2013.01.019

E-mail addresses: [email protected], mmm0010@t

Blogs and social networks have recently become a valuable resource for mining sentiments in fields asdiverse as customer relationship management, public opinion tracking and text filtering. In fact knowl-edge obtained from social networks such as Twitter and Facebook has been shown to be extremely valu-able to marketing research companies, public opinion organizations and other text mining entities.However, Web texts have been classified as noisy as they represent considerable problems both at thelexical and the syntactic levels. In this research we used a random sample of 3516 tweets to evaluate con-sumers’ sentiment towards well-known brands such as Nokia, T-Mobile, IBM, KLM and DHL. We used anexpert-predefined lexicon including around 6800 seed adjectives with known orientation to conduct theanalysis. Our results indicate a generally positive consumer sentiment towards several famous brands. Byusing both a qualitative and quantitative methodology to analyze brands’ tweets, this study adds breadthand depth to the debate over attitudes towards cosmopolitan brands.

� 2013 Elsevier Ltd. All rights reserved.

1. Introduction

Opinions expressed in social networks play a major role in influ-encing public opinion’s behavior across areas as diverse as buyingproducts, capturing the ‘‘pulse’’ of stock markets and voting for thepresident (Bai, 2011; Eirinaki, Pisal, & Singh, 2012). An opinion maybe regarded as a statement in which the opinion holder makes aspecific claim about a topic using a certain sentiment (Kim & Hovy,2004). Web-generated opinions in blogs and social networks haverecently become a valuable resource for mining user sentimentsfor the purpose of customer relationship management, public opin-ion tracking and text filtering (Zhang, Zeng, Li, Wang, & Zuo, 2009).Online opinions have been recently analyzed using sentiment anal-ysis (SA). This is basically a natural language processing (NLP)application that uses computational linguistics and text miningto identify text sentiment, typically as positive, neutral or negative.This technique is also known in the text mining literature as emo-tional polarity analysis (EPA), opinion mining, review mining, orappraisal extraction (Zagal, Tomuro, & Shepitsen, 2012). Thus, SAcan be regarded as an automated knowledge discovery techniquethat aims at finding hidden patterns in a large number of reviews,blogs or tweets. To calculate a sentiment score, the sentiment ob-tained from the text is compared to a lexicon or a dictionary todetermine the strength of the sentiment. For example, the lexicalresource SentiWord, which includes around 200,000 entries, usesa semi-supervised method to assign each word with positive,negative and objective scores. For instance, as Fig. 1 illustrates, a

ll rights reserved.

igermail.auburn.edu

negative word might have in one of its senses a sentiment scoreof negative 0.375, positive 0.125 and objective 0.5.

Knowledge obtained from social networks are extremely valu-able because millions of opinions expressed about a certain topicare highly unlikely to be biased. The affective nature of such opin-ions makes them easily understandable by the majority of readers,which increasingly make them the basis for making decisionsregarding marketing research, business intelligence, stock marketprediction and image monitoring (Montoyo, Martiniz-Barco, & Bal-ahur, Forthcoming). However, almost all online text-based com-munications ignore the rules of spelling and grammar. In fact,Web texts have been classified as noisy as they still pose consider-able problems both at the lexical and the syntactic levels (Boiy &Moens, 2009). At the lexical level, jargon, contractions of existingwords/abbreviations, the use of emoticons and the creation ofnew words are the norm. At the syntactic level, we can hardlyspeak of real sentences. This writing style is evident in most formsof computer-mediated communication forums such as social net-work sites, bulletin boards and chat rooms (e.g., Derks, Fischer, &Bos, 2008). Although language purists might argue that such ten-dency represents poor language use, Thelwall, Buckley, Paltoglou,Cai, and Kappas (2010) claim that such use is prompted by techno-logical advancements along with social factors. This complicatingfactor pertaining to informal Web texts’ sentiment detection hasbeen dealt with through several techniques, including word sensedisambiguation (Pederson, 2001), accurate detection of negation(Dave, Lawrence, & Pennock, 2003), and inferring semantic orienta-tion from association (Turney & Littman, 2003). Dealing success-fully with this problem has led to a plethora of online sentimentanalyses in texts written in languages as diverse as English

http://dx.doi.org/10.1016/j.eswa.2013.01.019

mailto:[email protected]

mailto:[email protected]

http://dx.doi.org/10.1016/j.eswa.2013.01.019

http://www.sciencedirect.com/science/journal/09574174

http://www.elsevier.com/locate/eswa

Fig. 1. Quantifying sentiment scores.

4242 M.M. Mostafa / Expert Systems with Applications 40 (2013) 4241–4251

(e.g., Jansen, Zhang, Sobel, & Chowdury, 2009), Chinese (e.g., Xu,Liao, & Li, 2008) Arabic (Ahmed & Almas, 2005), and multi-languages (Abbasi, Chen, & Salem, 2008).

Although several studies have recently investigated SA (e.g., Cai,Spangler, Chen, & Zhang, 2010; Leong, Lee, & Mak, 2012), no previ-ous studies have focused solely on investigating consumers’ senti-ments towards major worldwide brands such as IBM, Nokia andDHL. In this study we aim to fill this void. We believe that by inves-tigating brans’ polarity the study adds depth to the knowledge baseon text mining. By using both a qualitative and quantitative meth-odology to analyze brand comments, this study also adds breadthto the debate over brand quality as perceived by consumers. Final-ly, by focusing solely on online texts, rather than on traditional off-line data, this study enriches the knowledge base of this under-represented area. More specifically, this research aims attemptsto answer the following research questions:

RQ1. Can social networks’ opinion mining techniques be usedsuccessfully to detect hidden patterns in consumers’ sentimentstowards global brands?; and

RQ2. Can companies effectively use the blogosphere to redesigntheir marketing and advertising campaigns?

This paper is organized as follows. Next section provides a briefliterature review on the major areas of SA applications. Section 3deals with the method used to conduct the analysis. In this sectionissues related to research design and, sampling and data analysistechniques are presented. In Section 4, the results of sentimentanalysis are presented. Finally, Section 5 presents research implica-tions and limitations. This section also explores avenues for futureresearch.

2. Literature review

SA techniques have been recently utilized in applications suchas extracting suggestions from consumers’ product reviews (e.g.,Vishwanath & Aishwarya, 2011), classifying consumers’ positiveand negative product reviews (e.g., Turney, 2002), tracking senti-ment trends in online discussion boards (e.g., Tong, 2001), detect-ing Internet hot spots (e.g., Li & Wu, 2010), tracking politicalopinions (e.g., Thomas, Pang, & Lee, 2006), determining consumers’dissatisfaction with online advertising campaigns (e.g., Qiu et al.,2010), tracking emotions in emails (Mohammad, 2012), predictingstock market movements (e.g., Wong, Xia, Xu, Wu, & Li, 2008) anddifferentiating between informative and emotional social mediacontent (e.g., Denecke & Nejdi, 2009). An extensive literature re-view suggests that most SA applications might be classified into

four distinct categories: product reviews, movie reviews, politicalorientation extraction and stock market predictions.

2.1. Product reviews

Blair-Goldensohn et al. (2008) used Google Maps data as inputin order to analyze consumer sentiments towards hotels, depart-ment stores and restaurants. Using polarity values (positive/nega-tive), the system developed was able to summarize sentimentregarding different aspects of the service provided such as valuefor money and ambience. In the same vein, Yi, Nasukawa, Bunescu,and Niblack (2003) developed a sentiment analyzer to evaluateconsumers’ opinions regarding digital camera features. The systemused online text reviews to extract consumers’ sentiments regard-ing important features of digital cameras such as resolution andpicture quality. Liu, Huang, An, and Yu (2007) used probabilisticLatent Sentimen Analysis (PLSA) to predict future product salesby examining bloggers’ sentiment.

Feldman, Fresko, Netzer, and Ungar (2007) developed a polaritysystem to analyze consumers’ comparison comments posted ondiscussion boards. The system used online information such as‘‘300 �C Touring looks so much better than the Magnum’’ to ana-lyze consumers’ sentiments regarding several product aspects suchas style, noise, quality and price. Hu and Liu (2004) used machinelearning methods to extract and summarize consumers’ senti-ments related to several electronic products, including mp3 play-ers, digital cameras and mobile/cellular phones. The systemdeveloped classified each review into positive or negative opinionand predicted future buying behavior. In a similar study, Miyoshiand Nakagami (2007) analyzed consumer sentiments regardingelectronic products using adjective-noun pairs in a sentence.

In a recent study, Zhang, Xu, and Wan (2012) developed an ex-pert system, called Weakness Finder, in order to analyze consum-ers’ sentiments in Chinese language online texts. The systemextracts attitudes towards product features such as quality andprice based on a morpheme-based analysis. The system wastrained to utilize explicit and implicit sentiments to determineeach sentiment’s polarity regarding products’ weaknesses. Thisstudy extended previous work by Ding, Liu, and Yu (2008) andby liu (2010) because it took into consideration several linguisticaspects such as the adverbs of degree and the negation. Similarly,Abrahams, Jiao, Wang, and Fan (Forthcoming) employed text min-ing techniques to detect online consumer complaints regardingseveral automobile models. The authors found that consumer sen-timents may be used to categorize and prioritize vehicle defects.Pekar and Ou (2008) used sentiment analysis technique to evaluate268 reviews of major hotels based on customers’ reviews posted onthe website ‘‘epinions.com’’. The authors used attributes such asfood, room service, facilities and price to automatically analyzesentiments expressed towards those features. Finally, Na, Khoo,and Wu (2005) used support vector machines to classify 1800product reviews into either recommended/positive sentiment ornot recommended/negative sentiment. The authors used erroranalysis to improve initially obtained classification accuracy. Majorsources of error in classification were due to negation, superficialwords and comments on parts of the products.

2.2. Movie reviews

Na, Thet, and Khoo (2010) used a sample of 520 online moviereviews to conduct sentiment analysis. The authors compared tex-tual characteristics of consumers’ reviews across four differentgenres to investigate sentiments expressed towards movies suchas ‘‘Slumdog Millionaire’’, ‘‘American Gangster’’ and ‘‘Burn afterReading’’. Genres analyzed included discussion board threads, userreviews, critic reviews and bloggers’ postings. This study focused

Fig. 2. Top 20 countries in Twitter accounts as of January 1, 2012 (source: Semiocast.com).

Table 1Brands included in the study and their average sentiment scores.

ID Airline name # of tweets Mean sentiment score

1 IBM 825 0.25082 Nokia 600 0.53513 Pfizer 373 �0.40214 Lufthansa 239 0.23435 KLM 222 0.07176 DHL 203 0.25627 Comcast 200 �0.37018 Mobinil 123 0.79519 Citi group bank 115 �0.060810 Air India 108 �0.012811 Novartis 101 0.158412 T-Mobil 100 �0.331213 US Bank 98 �0.861714 Samsung 86 0.069715 Al-Jazeera English 66 0.073516 Egypt air 57 �0.1853

Total and grand mean 3516 0.0270

M.M. Mostafa / Expert Systems with Applications 40 (2013) 4241–4251 4243

on linguistic aspects of comments such as vocabulary, sentencelength and part-of-speech distribution. The authors found thatcomments on discussion boards and user reviews contain moreverbs and adverbs compared to the heavy usage of nouns and prep-ositions found in bloggers and critic postings. The study also iden-tified the most frequent positive and negative terms used acrossdifferent genres along with the distribution patterns of such terms.

Zhuang, Jing, and Zhu (2006) used machine learning methods tosummarize online texts movie reviews sentiments. The authorsaim was to find feature opinion pairs in consumers’ reviews bydetecting feature classes such as ‘‘sound effects’’ and the statedopinion such as ‘‘excellent.’’ In a similar research design, Pang,Lee, and Vaithyanathan (2002) used support vector machines(SVM) to classify online sentiment classification of movie reviews.

The authors used both single words (unigrams) and pairs of adja-cent words (bigrams) to conduct the analysis. Compared withother machine learning classification methods, the SVM techniqueachieved the highest accuracy (83% correct classification). Based onthe Page Rank algorithm, Wijaya and Bressan (2008) used onlineuser reviews to evaluate movies. The authors’ reported resultscompared favorably with the rankings reported by the box office.

Thet, Na, and Khoo (2008a) used machine learning andinformation extraction techniques such as pronoun resolutionand co-referencing to analyze sentiment orientation of moviereview online texts. The authors correctly segmented customers’reviews into relevant sections pertaining to different aspects ofthe movie such as the cast, the director and the overall rating. Ina second study, Thet, Na, and Khoo (2008b) proposed an automaticmethod for determining movie reviews’ sentiment orientation andstrength. The authors used a computational linguistics approachtaking into consideration the grammatical dependency structureof each clause analyzed.

2.3. Political orientation

Larsson and Moe (2011) investigated Twitters’ emotional polar-ity during the 2010 Swedish election using around 100,000 tweetsdealing with the election. The authors suggested a novel approachto classify high-end tweets messages among microbloggers intoseveral categories such as senders, receivers and sender-receivers.Similarly, Tumasjan, Sprenger, Sandner, and Welpe (2011) investi-gated 100,000 tweets message referring either to a politician or toa political party in Germany to predict election outcome. Williamsand Gulati (2008) also found that electoral success may be pre-dicted accurately by the total number of Facebook supporters.The authors also found that tweets sentiment analysis can be usedto accurately predict election outcome.

Table 2Sample of tweets for Air India.

Tweet ManualEvaluation

[1] ‘‘#Air India is so fab. nMaybe I under rated it all these years. Maybe it is gonna b my fav Indian airline now. nMaybe my life is about to change’’ Positive[2] ‘‘Irrespective of what people say, I still like #Air India, far better than snooty staff on most of these so called low cost carriers’’ Positive[3] ‘‘Horrible experience with #Air India – bag misplaced on AI101, no updates for two days. Is anyone there? Hello? Anyone with similar

experience?’’Negative

[4] ‘‘Don’t fly AI then! RT @Terrell_Raupp @ikaveri Air India gives the worst flight experience to its customer http://t.co/NM0abRNc #Air India’’ Negative[6] ‘‘Sat on the tarmac for the last hour. Thanks #Air India. Grow a spine and push back to the babu sitting in air traffic control:-/’’ Negative[7] ‘‘#Air India is regressing to the babudom days of yore. The #airline does not respect its staff & in turn the staff does not respect customers’’ Negative[8] ‘‘RT @vageee: #Indian Hockey is like #Air India, all we can do is think about its past glory days’’ Negative[9] ‘‘It takes half an hour to get a ticket from #Air India counter. They are definately nt living in jet age’’ Negative[10] ‘Flew #Air India after a while. Took off and landed on time, good in-flight service, bags were on the carousel in 5 min. Impressed!’’ Positive[11] ‘‘#Air India express flight schedule changes effective Wednesday http://t.co/Z7JbwsZD’’ Neutral[12] ‘‘ @JourneyMart #Travel #Air India 4 days, 3 flights, all on AI, all ON TIME. Superb service although no tender touches! Gr8 seat pitch. WOW’’ Positive

Table 3Word frequencies of a sample of tweets after removing stopping and unique words.

Freq %Shown %Processed %Total

GLOBAL 41 4.90% 0.80% 0.60%CENTER 38 4.50% 0.80% 0.50%DEVELOPMENT 37 4.40% 0.70% 0.50%FLIGHT 35 4.20% 0.70% 0.50%EGYPT 34 4.10% 0.70% 0.50%FELLOW 28 3.30% 0.60% 0.40%FLIGHTS 27 3.20% 0.50% 0.40%PRICE 21 2.50% 0.40% 0.30%QATAR 20 2.40% 0.40% 0.30%AIRLINES 19 2.30% 0.40% 0.30%DAY 19 2.30% 0.40% 0.30%BOARD 18 2.20% 0.40% 0.20%BOOK 18 2.20% 0.40% 0.20%BUILDING 18 2.20% 0.40% 0.20%HOUSE 18 2.20% 0.40% 0.20%SYRIA 18 2.20% 0.40% 0.20%DOUBLED 17 2.00% 0.30% 0.20%FIGHTING 17 2.00% 0.30% 0.20%FLEE 17 2.00% 0.30% 0.20%FRIENDS 17 2.00% 0.30% 0.20%HALL 17 2.00% 0.30% 0.20%MONOPOLY 17 2.00% 0.30% 0.20%RESIDENT 17 2.00% 0.30% 0.20%TICKETS 17 2.00% 0.30% 0.20%AVIATION 15 1.80% 0.30% 0.20%DOHA 15 1.80% 0.30% 0.20%DUBAI 15 1.80% 0.30% 0.20%FLYEGYPTAIR 15 1.80% 0.30% 0.20%CGD 14 1.70% 0.30% 0.20%INDIA 14 1.70% 0.30% 0.20%TURKISH 14 1.70% 0.30% 0.20%AIRLINE 12 1.40% 0.20% 0.20%BA 12 1.40% 0.20% 0.20%DREAMLINER 12 1.40% 0.20% 0.20%FLYING 12 1.40% 0.20% 0.20%TIME 12 1.40% 0.20% 0.20%VISITING 12 1.40% 0.20% 0.20%BEEF 11 1.30% 0.20% 0.10%CHICKEN 11 1.30% 0.20% 0.10%MUSCAT 11 1.30% 0.20% 0.10%QUESTION 11 1.30% 0.20% 0.10%REPRESENTS 11 1.30% 0.20% 0.10%SICK 11 1.30% 0.20% 0.10%AIRPORT 10 1.20% 0.20% 0.10%FLY 10 1.20% 0.20% 0.10%SERVICE 10 1.20% 0.20% 0.10%


In their seminal study on informal political texts, Malouf andMullen (2008) argued that SA is a useful technique that might beused in order to analyze possible ideological biases, politicalopinions and political judgments’ favorability. The authors usedthis technique to investigate political orientation among theusers of a specific US web site dedicated to political discussions.This study is important because it is probably the first to inves-tigate sentiment analysis in informal online political discussions-an area that is fast ‘‘becoming an important feature of theintellectual landscape of the Internet’’ (p. 179). Golbeck, Grimes,and Rogers (2010) also used SA to classify 6000 political tweetsby US Congress’ members. The authors found that the majorreason behind tweeting was to disseminate useful information,followed by tweets related to personal daily activities. Theauthors labeled the latter as a ‘‘vehicle for self-promotion’’(p. 1620). Similarly, Ekdale, Namkoong, and Perlmutter (2010)empirically analyzed political bloggers’ behavior in the US. Theauthors found that the main reason for blogging was promptedby extrinsic motivations such as influencing public opinion orusing the blogosphere as a plausible alternative to traditionalmedia. In a study investigating microblogging participation with-in political environment, Gil De Zuniga, Puig-I-Abril, and Rojas(2009) found that the major reasons for blogging are basicallyextrinsic motivations.

Heavy tweets used by Egyptian protesters from January 25 toFebruary 11, 2011, which led ultimately to the forced resignationof ex-dictator Hosni Mubarak were extensively investigated (e.g.,Lim, 2012; Papacharissi & Oliveira, 2012). These studies foundthat Twitter was used by protesters as an alternative to theblocked access to the Internet. The continuous stream of eventsprovided by Twitter users was also found to be an accurate pre-dictor of the crisis outcome. Park, Lim, Sams, Nam, and Park(2011) analyzed 2000 comments posted on ten Korean politicians’visitor boards. The authors started by classifying the comments aspositive, negative, or irrelevant. The authors found that positivecomments represent the majority of all comments with a 51.3%.Negative comments represented 20.8% while irrelevant commentsrepresented 27.9%. In terms of gender, the authors founded thatfemale comments were associated with more positive commentscompared to male users (75.5% vs. 67.3%). Zappavigna (2011)used a large corpus of tweets (45,000) posted immediatelyafter Obama’s US presidential elections victory in 2008. Usingcomputational linguistics techniques, the author showed thatthe hashtag has ‘‘extended its meaning potential to operate as alinguistic marker referencing the target of evaluation in a tweet’’(p. 788). Other studies investigating political sentiment analysisinclude studies by Efron (2004), Thomas et al. (2006), Park, Kim,and Barnett (2004) and Sobkowicz, Kaschesky, and Bouchard(Forthcoming).

2.4. Stock market prediction

Using automated natural language processing and machinelearning techniques, Das and Chen (2001) classified sentiments ex-pressed on Yahoo! Finance’s discussion board. The authors re-ported 62% accuracy in classifying posts into positive sentiment,

Fig. 3. Proximity plot based on Egypt Air tweets.


negative sentiment or neutral/irrelevant sentiment. In a similarstudy, Gu, Konana, Liu, Rajagopalan, and Ghosh (2006) used com-ments posted on Yahoo! Finance’s discussion board to predict dif-ferent stocks’ future returns. Each post was classified into fivepossible categories: (2) for ‘‘strong buy’’; (1) for ‘‘buy’’; (0) for‘‘hold’’; (�1) for ‘‘sell’’; and (�2) for ‘‘strong sell.’’ In this studythe authors also used a weighting scheme to assign weights foreach sentiment obtained based on the reputation and previousaccuracy of the poster. Using a simulated environment to mimicreal trading, the authors reported around four percent increase inreturns over one month based on sentiment analysis.

Bollen, Mao, and Zeng (2011) found that the aggregation of mil-lions of tweets posted daily on Twitter can be used to predict stockmarket over time. The authors used measures such as daily Twitterposts over around ten months to predict the Dow Jones Industrialaverage closing values. To cross-validate the results, the authorsalso used the resulting time series of Twitter moods to detect thegeneral public’s response towards the outcome of the US presiden-tial campaign. Other studies investigated the relationship betweeninvestors’ sentiments and other factors such as stock returns fol-lowing a major earthquake (Shan & Gong, 2012), air disastersinvolving US vs. foreign airlines (Kaplanski & Levy, 2010) and localsports events (Chang, Chen, Chou, & Lin, 2012).

3. Method

3.1. Twitter sampling

Twitter is a microblogging service that was launched formallyon July 13, 2006. Unlike other social media, Twitter is considered

a microblog because its central activity revolves around postingshort updates or tweets using the Web or mobile/cell phones.The maximum size of the blog is 140 characters-roughly the sizeof a newspaper headline. According to Semiocast.com (2012), amarketing research company, there are now around 500 millionactive twitters. Fig. 2 shows top ranked countries according to ac-tive tweets in 2012 (Semiocast.com, 2012). Tweets are availablepublicly as a default, and are also directly broadcasted to the user’sfollowers (Bliss, Klouman, Harris, Danforth, & Dodds, 2012).

A recent analysis of Twitter activities found that more than 80%of the users either update their followers on what they actuallydoing or disseminate information regarding their daily experiences(Thelwall, Buckley, & Paltoglou, 2011). Since Twitter is the mostlarge, popular and well-known microblog Web site, it was selectedto conduct the analysis reported in this study. The data used repre-sent a random set of Twitter posts from July 18, 2012, to August 17,2012. The data comprised 3516 tweets for sixteen brands. To guar-antee representativeness, sample selection has been varied by dayof the week and hours in the day. Our sample size is comparable insize to Qiu, He, Zhang, Shi, Bu, and Chen’s (2010) sample, which in-cluded 3783 opinion sentences. Table 1 shows the random sampleof tweets for each brand included in the study. Following Thelwallet al. (2011), only tweets in English was chosen in order to removecomplications that might arise with analyzing multilingual tweets.

Table 2 shows a sample of tweets for Air India with a manualclassification of customers’ sentiments. As can be seen from the ta-ble, tweets represent a very noisy environment in which messagesposted to virtual audience includes abbreviated words, the @ andthe hashtag (#) characters, and heteroglossia-referring to othervoices in the tweets in order to convey interpersonal and ideational

Fig. 4. A 3-D map with link strengths and base lines shown.


meanings (Bakhtin, 1981). Huang, Thornton, and Efthimiadis(2010) found that the hashtag was invented by Twitter users earlyin 2008 to help followers find a specific tweet or post. As opposedto the hashtag, the @ character has been introduced to address atweet to another follower, which allows Twitter to function effec-tively as a collaboration and conversation system (Honeycutt &Herring, 2009).

3.2. Lexicon

Categorizing words for SA is a major step in applying the tech-nique. Broadly speaking, there are two widely used methods forsentiment orientation identification: the lexicon-based approachand the corpus-based method (Miao, Li, & Zeng, 2010). However,since the corpus-based method has rarely been used to analyzesentiment orientation, we will focus here on the lexicon-basedmethod. Nevertheless, both methods require a pre-defined dictio-nary or corpus of subjective words. The sentiment is determinedby comparing tweets against the expert-defined entry in the dic-tionary, which makes it easy to determine the polarity of a specificsentence. Thus, it is crucial to have an accurate classifier to be usedto construct indicators of sentiment. Previous research has typi-cally incorporated lexicons such as the manually coded General In-quirer (Stone, Dunphy, Smith, & Ogilvie, 1966), which includes over11000 hand-coded word stems in 182 categories, the LIWC dictio-

nary (Pennebaker, Mehl, & Niederhoffer, 2003), the SentiWordNet(Baccianella, Esuli, & Sebastiani, 2010), the Q-WordNet (Agerri &Garcia-Serrano, 2010) or the lexicon of subjectivity clues (Wiebe,Wilson, Bruce, Bell, & Martin, 2004). Automatically-coded lexiconshave recently been developed, including the sentiment-based lex-icon (Taboada, Brooke, Tofiloski, Voll, & Stede, 2011).

In this study we used the Hu and Liu (2004) lexicon to conductthe analysis because this dictionary has been used successfully in asimilar application (Miner et al., 2012). This lexicon includesaround 6800 seed adjectives with known orientation (2006 posi-tive words and 4783 negative words). The lexicon has recentlybeen updated by adding words based on a thorough search inthe WordNet. The lexicon size is similar to the recently used Opin-ion Finder lexicon, which included 2718 positive words and 4912negative words (Bollen et al., 2011). Our lexicon is also comparableto the dictionary used by Hu, Bose, Koh, and Liu (2012), which in-cludes 1635 positive words and 2005 negative words. Modifica-tions on this approach include the handling of negation (Das &Chen, 2001) and the weight of enhancer/intensifier words suchas the use of the ‘‘really’’ or ‘‘absolutely’’ (Turney, 2002). Other at-tempts have been recently tried to develop lexicons capable ofdetecting deep sentiments expressed in discussions among severalactors in informal situations (Maks & Vossen, 2012). Similarly,Reyes and Rosso (2012) developed a corpus to detect irony in con-sumers’ reviews.

Fig. 5. Sentiment scores for Nokia (top) and Pfizer (bottom). X-axis represents scoredistributions, Y-axis represents count/frequencies.


4. Results

4.1. Exploratory data analysis and visualizations

To conduct the qualitative part of this study we used QDA Miner4.0 software package (Provalis Research, 2011) for coding textualdata posted on Twitter. This software was selected because of itsextensive exploratory tools that can be used to identify hidden pat-terns in textual data. In order to analyze consumer sentiments to-wards brands, we started by generating relative frequency wordcounts. Table 3 shows the percentage of words in a random setof tweets.

From Table 3 we can see that words such as ‘‘global’’, ‘‘flight’’and ‘‘price’’ have the highest frequency for a brand such as EgyptAir. However, references were also made in the tweets to countriessuch as Syria, probably because of the ongoing uprising in thatcountry. Analyzing frequency of appearance or simply the incre-mental count of appearance of particular words or phrases mightprovide insights into a particular topic. In fact O’Leary (2011) ar-gues that despite the simplicity of such approach, it can be usedto predict characteristics of the topic analyzed. Fig. 3 shows a prox-imity plot constructed based on Egypt Air tweets. This figure showsvisually, on a single axis, the distance from a particular object to allother objects. This graph was constructed to extract huge amountof data based on a distance matrix. From this graph we see thatmost tweets were concerned with things like ‘‘price’’, ‘‘tickets’’and ‘‘monopoly.’’ However, Twitters seem to be also concernedwith fighting in neighboring Syria.

Fig. 4 shows a 3-D concept map constructed based on multidi-mensional scaling (MDS) technique. In this graph the closer thecases, the higher the tendency of co-occurrence and vice versa.The lines on the map represent levels of association among words.From the graph we can reconstruct the most influential tweets. Forexample, for Egypt Air, we can see the following pattern: closenessto home as represented by words such as ‘‘Egypt’’, ‘‘Dubai’’ and‘‘Resident’’, personification as represented by names of peopletweeting, and relevance and significance as represented by wordssuch as ‘‘Dreamliner’’, ‘‘Service’’, and ‘‘Flight.’’ This result is in linewith a recent study using centering resonance analysis, a computa-tional discourse analysis technique, on a random sample of 9000Egyptian tweets (Papacharissi & Oliveira, 2012).

4.2. Overall sentiment scores

We used the twitteR, the plyr, stringr and the ggplot2 librariesin the R software package version 2.15 to conduct the quantitativesentiment score. Fig. 5 shows the distribution of sentiment scoresobtained for Nokia and Pfizer (similar graphs were constructed forremaining brands). From the graph, we immediately recognizesome asymmetry. For example, the bars at +1 are much larger forNokia compared to the +1 bars for Pfizer. It is also evident thatthe bars at �1 are much larger for Pfizer compared to the �1 barsfor Nokia. This makes it clear that the overall sentiment score forNokia is generally better than the sentiment score for Pfizer.

The visualization of the sentiment distribution in Fig. 5 furtherunderlines the fact that most tweets fall either on the neutral point(0) or within the band of circa �1/+1, which is an indication thatseveral tweets are not very affective. Although this result is in linewith Lindgren (2012), we can focus only on positive or negativesentiments. Following Miner et al. (2012), we ignored the middleand constructed sentiment scores for a random sample of only po-sitive and negative sentiments. Fig. 6 shows results for only threebrands. From Fig. 6, we clearly see that most tweet comments werepositive for both Lufthansa and DHL. However, most of the tweetswere negative for T-Mobile.

Finally, we used the StreamGraph software package (Clark,2008) to visualize the trend of tweets across a period of time forall brands. Fig. 7 shows tweets trend for three brands: Nokia, Pfizerand Al-Jazeera English. This graph is useful since it represents mul-tiple time series data stacked one on top of the other (Havre, Het-zler, Whitney, & Nowell, 2002). Since the total frequency of allfeatures represents the height of the curve, each time series datashould be read off the figure not as the cumulative height butrather as starting with zero. As can be seen the graph is character-ized by a number of spikes, indicating an increase in tweets’ fre-quency at those particular times. Interestingly, we see spikes onthe top part of the figure representing Nokia’s introduction of thenew smart phone Nokia Lumia brand.

5. Implications, limitations and future research

In this study we analyzed sentiment polarity of more than3500 social media tweets expressing attitudes towards sixteenglobal brands. Social media users represent 67% of around a bil-lion Internet active users (Eirinaki et al., 2012). Although a singletweet is limited to 140 characters in length, the millions of tweetsposted on Twitter almost on a daily basis might provide an unbi-ased representation of consumers’ sentiment towards servicesand brands. Capturing consumers’ opinions and gaining knowl-edge about consumer preferences has long been a major concernfor marketing researchers. However, traditional marketing meth-ods such as focus groups and face-to-face interviewing are bothcostly and time consuming. In contrast, tweets and blogs arereadily available for free. Such consumer generated media are alsofree of bias that might be introduced by the interviewer in case of

Fig. 6. Sentiment analysis for a random tweets sample-after eliminating neutral tweets-for Lufthansa (top), DHL (middle) and T-Mobile (bottom) brands.


personal interviews. Moreover, consumers opinion-based ex-pressed qualitatively may easily be benchmarked against objec-tive measures such as sales data, revenues, or stock price. Thus,companies can utilize such online textual content in an effort togain insight into consumers’ opinions regarding available prod-ucts and services. Ignoring consumer generated sentiments mightput companies in a competitive disadvantage and could also cre-ate significant brand image problems. The speed of social mediamight also render companies’ advertising and publicity using tra-ditional media useless.

Based on the fact that around 20% of microblogs mention a brandname (Jansen et al., 2009), we argue that managing brand perceptionon Twitter and other social media should form part of the company’soverall proactive marketing strategy. Maintaining a constant pres-ence on such media channels should also be an important part ofthe company’s branding and advertising campaigns. Companiescan use the blogosphere in a smart way to disseminate informationneeded by its customers and to monitor Twitters and bloggers’ dis-cussions regarding its brand. By doing this, companies can tracktweets and intervene immediately to communicate with dissatisfiedcustomers. On the other hand, advertising campaigns might makeuse of positive tweets, which can form a part of the company’s viralmarketing efforts. Thus, companies may use consumers’ tweets as afeedback about services and products by encouraging electronicword of mouth (e-WOM). This can be done online without investinghuge amounts on traditional advertising and marketing campaigns.On the other hand, companies should not ignore negative tweets be-cause such tweets might be used to detect what is not going rightwith a product or a service. Ultimately, tweets can be used effectivelyto identify consumers’ preferences, to detect dissatisfaction relatedto a product defect, and to correct unintended errors. Since tweetsenable companies to be more efficient and functional in dealing with

customers, they should incorporate SA into their text retrieval tech-nologies and into their search engines. This is because SA can be ex-tremely useful in areas such as analyzing consumer trends, handlingcustomers’ feedback and targeting advertising campaigns.

However, it should be noted that while we conducted SA toobjectively classify consumers’ opinions, our analysis does not re-veal the underlying reason behind forming such opinions. Futureresearch using sentiment topic recognition (STR) should be con-ducted to determine the most representative topics discussed be-hind each sentiment. Through this analysis, it should be possibleto gain overall knowledge regarding the underlying causes of posi-tive or negative sentiments. It should also be noted that while thelexicon-based approach used in this study can detect basic senti-ments, such approach may sometimes fall short of recognizingthe subtle forms of linguistic expression used in situations suchas sarcasm, irony or provocation. For instance, Boiy and Moens(2009) provide an excellent example by showing that the part ofthe sentence that follows ‘‘même si/even if’’ in a French sentenceexpresses the least affective feeling ‘‘[même si le film a eu beau-coup de succès, je le trouvais vraiment nul!/even though the moviehad a lot of success, I really found it nothing!]’’. Future researchshould attempt to find a way to deal with this problem. A huge cor-pus that includes large training data sets representing such idio-matic usage may be worth trying. Finally, opinions expressed byconsumers might in fact be a manipulation of online vendors’ opin-ions posing as real consumers. This manipulation might distortsentiments of real consumers. Future research should also attemptto detect genuine sentiments from opinions that merely reflect theposition of vendors interested in selling more products or services.

Despite these limitations, we believe that this study contributesto the existing literature in text mining and consumer behavior.First, we used a number of well-known brands, which ensures that

Fig. 7. Twitter stream graph (1000 tweets each) for Nokia (top), Pfizer (middle) and Al-Jazeera English brands (bottom).


our findings are practical, influential and generalizable. Second, weapproached our analysis using the most widely-used microblog-ging site-Twitter by employing a mixed methods approach basedon both qualitative and quantitative methods. This ensures our re-

sults robustness. Finally, by focusing only on consumer tweets wecontribute to the growing body of literature on e-WOM (e.g.,Jansen et al., 2009). Therefore, we believe that our research istimely and impactful.


References

Abbasi, H., Chen, H., & Salem, A. (2008). Sentiment analysis in multiple languages:Feature selection for opinion classification in web forums. ACM Transactions onInformation and Systems, 26, 1–34.

Abrahams, A., Jiao, J., Wang, G., & Fan, W. (Forthcoming). Vehicle defect discoveryfrom social media. Decision Support Systems.

Agerri, R., & Garcia-Serrano, A. (2010). Q-WordNet: Extracting polarity fromWordNet senses. In Paper presented at the 7th conference on internationallanguage resources and evaluation. Malta (www.irec-conf.org).

Ahmed, K., & Almas, Y. (2005). Visualising sentiments in financial texts? InProceedings of the 9th international conference on, information visualization (pp.363–368).

Baccianella, S., Esuli, A., & Sebastiani, F. (2010). SentiWordNet 3.0: An enhancedlexical resource for sentiment analysis and opinion mining. In Proceedings of the7th conference on international, language resources and evaluation (www.irec-conf.org).

Bai, X. (2011). Predicting consumer sentiments from online text. Decision SupportSystems, 50, 732–742.

Bakhtin, M. (1981). The dialogic imagination. Austin: University of Texas Press.Blair-Goldensohn, S., Hannan, K., McDonald, R., Neylon, T., Reis, G., & Reynar, J.

(2008). Building a sentiment summarizer for local service reviews. In paperpresented at the www 2008 workshop on NLP challenges in the informationexplosion era (NLPIX 2008), Beijing, April, 22.

Bliss, C., Klouman, I., Harris, K., Danforth, C., & Dodds, P. (2012). Twitter reciprocalnetworks exhibit assortativity with respect to happiness. Journal ofComputational Science, 3, 388–397.

Boiy, E., & Moens, M. (2009). A machine learning approach to sentiment analysis inmultilingual web texts. Information Retrieval, 12, 526–558.

Bollen, J., Mao, H., & Zeng, X. (2011). Twitter mood predicts the stock market. Journalof Computational Science, 2, 1–8.

Cai, K., Spangler, S., Chen, Y., & Zhang, L. (2010). Leveraging sentiment analysis fortopic detection. Web Intelligence and Agent Systems, 8, 291–302.

Chang, S., Chen, S., Chou, R., & Lin, Y. (2012). Local sports sentiment and returns oflocally headquartered stocks: A firm-level analysis. Journal of Empirical Finance,19, 309–318.

Clark, J. (2008). Twitter topic stream (www.neoformix.com).Das, S., & Chen, M. (2001). Yahoo! For Amazon: Sentiment parsing from small talk

on the com. European Finance Association Meeting: Barcelona.Dave, K., Lawrence, S., & Pennock, D. (2003). Mining the peanut gallery: Opinion

extraction and semantic classification of product reviews. Proceedings of the 12thinternational conference on the world wide web (WWW-03). New York: ACMPress. 519–528.

Denecke, K., & Nejdi, W. (2009). How valuable is medical social media data? Contentanalysis of the medical web. Information Sciences, 179, 1870–1880.

Derks, D., Fischer, A., & Bos, A. (2008). The role of emotion in computer-mediatedcommunication: A review. Computers in Human Behavior, 24, 766–785.

Ding, X., Liu, B., & Yu, P. (2008). A holistic lexicon-based approach to opinion mining.In Proceedings of the international conference on web search and web data mining(pp. 231–240). CAN.

Efron, M. (2004). Cultural orientation: Classifying subjective documents by co-citation analysis. AAAI Fall Symposium on Style and Meaning in Language, Art,and Music.

Eirinaki, M., Pisal, S., & Singh, J. (2012). Feature-based opinion mining and ranking.Journal of Computer and System Sciences, 78, 1175–1184.

Ekdale, B., Namkoong, K., & Perlmutter, D. (2010). Why blog? (Then and now):Exploring the motivation for blogging by popular American political bloggers.New Media and Society, 12, 217–234.

Feldman, R., Fresko, M., Netzer, P., & Ungar, L. (2007). Extracting productcomparisons from discussion boards. In Proceedings of the 7th IEEEinternational conference on data mining (ICDM’ 2007) (pp. 469–474). LosAlamitos: CA.

Gil De Zuniga, H., Puig-I-Abril, E., & Rojas, H. (2009). Weblogs, traditional sourcesonline and political participation: An assessment of how the internet ischanging the political environment. New Media and Society, 11, 553–574.

Golbeck, J., Grimes, J., & Rogers, A. (2010). Twitter use by the US congress.Journal of the American Society for Information Science and Technology, 61,1612–1621.

Gu, B., Konana, P., Liu, A., Rajagopalan, B., & Ghosh, J. (2006). Predictive value ofstock message board sentiments. McCombs Research Paper Series No. IROM-11-06.

Havre, S., Hetzler, E., Whitney, P., & Nowell, L. (2002). ThemeRiver: Visualizingthematic changes in large document collections. IEEE Transactions onVisualization and Computer Graphics, 8, 9–20.

Honeycutt, C., & Herring, S. (2009). Beyond microblogging: Conversation andcollaboration via Twitter. In Proceedings of the 42nd Hawaii internationalconference on system sciences (HICSS 09) (pp. 1–10).

Hu, M., & Liu, B. (2004). Mining and summarizing customer reviews. In Proceedingsof the 10th international conference on knowledge discovery and data mining (ACMSIGKDD 2004) (pp. 168–177).

Hu, N., Bose, I., Koh, N., & Liu, L. (2012). Manipulation of online reviews: An analysisof ratings, readability, and sentiments. Decision Support Systems, 52, 674–684.

Huang, J., Thornton, K., & Efthimiadis, E. (2010). Conversational tagging in Twitter.Proceedings of the 21st ACM conference on hypertext and hypermedia. New York:ACM Press, pp. 173–178.

Jansen, B., Zhang, M., Sobel, K., & Chowdury, A. (2009). Twitter power: Tweets aselectronic word of mouth. Journal of the American Society for Information Scienceand Technology, 60, 2169–2188.

Kaplanski, G., & Levy, H. (2010). Sentiment and stock prices: The case of disasters.Journal of Financial Economics, 95, 174–201.

Kim, S., & Hovy, E. (2004). Determining the sentiment of opinions. In Proceedings ofthe international conference on computational linguistics (COLING 2004) EastStroudsburg, PA, 1367.

Larsson, A., & Moe, H. (2011). Studying political microblogging: Twitter users in the2010 Swedish election campaign. New Media and Society, 14, 727–747.

Leong, C., Lee, Y., & Mak, W. (2012). Mining sentiments in SMS texts for teachingevaluation. Expert Systems with Applications, 39, 2584–2589.

Li, N., & Wu, D. (2010). Using text mining and sentiment analysis for online forumshotspot detection and forecast. Decision Support System, 48, 354–368.

Lim, M. (2012). Clicks, cabs, and coffee houses: Social media and oppositionalmovements in Egypt, 2004–2011. Journal of Communication, 62, 231–248.

Lindgren, S. (2012). It took me about half an hour, but I did it! Media circuits andaffinity spaces around how-to videos on YouTube. European Journal ofCommunication, 27, 152–170.

Liu, B. (2010). Sentiment analysis and subjectivity. Handbook of Natural LanguageProcessing.

Liu, Y., Huang, X., An, A., & Yu, X. (2007). ARSA: A sentiment-aware model forpredicting sales performance using blogs. In Proceedings of the 30th annualinternational ACM SIGIR conference on research and development in informationretrieval (pp. 607–614). New York.

Maks, I., & Vossen, P. (2012). A lexicon model for deep sentiment analysis andopinion mining applications. Decision Support Systems, 53, 680–688.

Malouf, R., & Mullen, T. (2008). Taking sides: User classification for informal onlinepolitical discourse. Internet Research, 18, 177–190.

Miao, Q., Li, Q., & Zeng, D. (2010). Fine-grained opinion mining by integratingmultiple review sources. Journal of the American Society for Information Scienceand Technology, 61, 2288–2299.

Miner, G., Delen, D., Elder, J., Fast, A., Hill, T., & Nisbet, B. (2012). Practical text miningand statistical analysis for non-structured text data applications. Amsterdam, TheNetherlands: Academic Press.

Miyoshi, T., & Nakagami, Y. (2007). Sentiment classification of customer reviews onelectronic products. In Proceedings of the international conference on systems,man and cybernetics (2028–2033).

Mohammad, S. (2012). From once upon a time to happily ever after: Trackingemotions in mail and books. Decision Support Systems, 53, 730–741.

Montoyo, A., Martiniz-Barco, P., & Balahur, A. (Forthcoming). Subjectivity andsentiment analysis: An overview of the current state of the area and envisageddevelopments. Decision Support Systems.

Na, J., Khoo, C., & Wu, P. (2005). Use of negation phrases in automatic sentimentclassification of product reviews. Library Collections, Acquisitions, and TechnicalServices, 29, 180–191.

Na, J., Thet, T., & Khoo, C. (2010). Comparing sentiment expression in movie reviewsfrom four online genres. Online Information Review, 34, 317–338.

O’Leary, D. (2011). Blog mining-review and extensions: From each according to hisopinion. Decision Support Systems, 51, 821–830.

Pang, B., Lee, L., Vaithyanathan, S. (2002). Thumbs up? Sentiment classificationusing machine learning techniques. In Proceedings of the 2002 conference onengineering methods in natural language processing (pp. 79–86). Morristown, NJ.

Papacharissi, Z., & Oliveira, M. (2012). Affective news and networked publics: Therhythms of news storytelling on #Egypt. Journal of Communication, 62,266–282.

Park, H., Kim, C., & Barnett, G. (2004). Socio-communicational structure amongpolitical actors on the web. New Media and Society, 6, 403–423.

Park, S., Lim, Y., Sams, S., Nam, S., & Park, H. (2011). Networked politics on cyworld:The text and sentiment of Korean political profiles. Social Science ComputerReview, 29, 288–299.

Pederson, T. (2001). A decision tree of bigrams is an accurate predictor of wordsense. In Proceedings of the second annual meeting of the North Americanchapter of the association for computational linguistics (pp. 79–86).

Pekar, V., & Ou, S. (2008). Discovery of subjective evaluations of product features inhotel reviews. Journal of Vacation Marketing, 14, 145–155.

Pennebaker, J., Mehl, M., & Niederhoffer, K. (2003). Psychological aspects of naturallanguage use: Our words, our selves. Annual Review of Psychology, 54, 547–577.

Provalis Research (2011). QDA Miner version 4.0 User Manual. Montreal, QC,Canada.

Qiu, G., He, X., Zhang, F., Shi, Y., Bu, J., & Chen, C. (2010). DASA: Dissatisfaction-oriented advertising based on sentiment analysis. Expert Systems withApplications, 37, 6182–6191.

Reyes, A., & Rosso, P. (2012). Making objective decisions from subjective data:Detecting irony in customer reviews. Decision Support Systems, 53, 754–760.

Semiocast.com (2012). Geolocation analysis of Twitter accounts (accessed onAugust 5, 2012).

Shan, L., & Gong, S. (2012). Investor sentiment and stock returns: Wenchuanearthquake. Finance Research Letters, 9, 36–47.

Sobkowicz, P., Kaschesky, M., & Bouchard, G. (Forthcoming). Opinion mining insocial media: Modeling, simulation, and forecasting political opinions in theweb. Government Information Quarterly.

Stone, P., Dunphy, D., Smith, M., & Ogilvie, D. (1966). The general inquirer: A computerapproach to content analysis. Cambridge, MA: The MIT Press.

Taboada, M., Brooke, J., Tofiloski, M., Voll, K., & Stede, M. (2011). Lexicon-basedmethods for sentiment analysis. Computational Linguistics, 37, 267–307.


Thelwall, M., Buckley, K., & Paltoglou, G. (2011). Sentiment in Twitter events. Journalof the American Society for Information Science and Technology, 62, 406–418.

Thelwall, M., Buckley, K., Paltoglou, G., Cai, D., & Kappas, A. (2010). Sentimentstrength detection in short informal text. Journal of the American Society forInformation Science and Technology, 61, 2544–2558.

Thet, T., Na, J., & Khoo, C. (2008a). Sentiment classification of movie reviews usingmultiple perspectives. Proceedings of the international conference on Asian digitallibraries (ICADL). Berlin: Springer Verlag. 184–193.

Thet, T., Na, J., & Khoo, C. (2008b). Aspects-based sentiment analysis of moviereviews on discussion boards. Journal of Information Science, 36, 823–848.

Thomas, M., Pang, B., & Lee, L. (2006). Get out the vote: Determining support oropposition from congressional floor-debate transcripts. In Proceedings of the2006 conference on empirical methods in natural, language processing (EMNLP).

Tong, R. (2001). An operational system for detecting and tracking opinions in onlinediscussions. In Working notes of the workshop on operational text classification(pp. 1–6). New Orleans, LA.

Tumasjan, A., Sprenger, T., Sandner, P., & Welpe, I. (2011). Election forecasts withTwitter: How 140 characters reflect the political landscape. Social ScienceComputer Review, 29, 402–418.

Turney, P. (2002). Thumbs up or thumbs down? Semantic orientation applied tounsupervised classification of reviews. In Proceedings of the 40th annual meetingof the association for computational linguistics (ACL 02) (pp. 417–424).Philadelphia, PA.

Turney, P., & Littman, M. (2003). Measuring praise and criticism: Inference ofsemantic orientation from association. ACM Transactions on Information Systems,21, 315–346.

Vishwanath, J., & Aishwarya, S. (2011). User suggestions extraction from customerreviews. International Journal on Computer Science and Engineering, 3,1203–1206.

Wiebe, J., Wilson, T. T., Bruce, R., Bell, M., & Martin, M. (2004). Learning subjectivelanguage. Computational Linguistics, 30, 277–308.

Wijaya, P., & Bressan, S. (2008). A random walk on the red carpet: Rating movies withuser reviews and Page Rank. Proceedings of the 17th ACM conference oninformation and knowledge management. New York: ACM Press. 951–960.

Williams, C., & Gulati, G. (2008). What is a social network worth? Facebook and voteshare in the 2008 presidential primaries. In The annual meeting of the Americanpolitical science association (pp. 1–17). Boston, MA: APSA.

Wong, K., Xia, Y., Xu, R., Wu, M., & Li, W. (2008). Pattern-based opinion mining forstock market trend prediction. International Journal of Computer processing ofLanguages, 21, 347–361.

Xu, D., Liao, S., & Li, Q. (2008). Combining empirical experimentation and modelingtechniques: A design research approach for personalized mobile advertisingapplications. Decision Support Systems, 44, 710–724.

Yi, J., Nasukawa, T., Bunescu, R., & Niblack, W. (2003). Sentiment analyzer:Extracting sentiments about a given topic using natural language-processingtechniques. In Proceedings of the 3rd IEEE international conference on data mining(ICDM’ 2003) (pp. 427–434). Los Alamitos, CA.

Zagal, J., Tomuro, N., & Shepitsen, A. (2012). Natural language processing in gamestudies research: An overview. Simulation and Gaming, 43, 356–373.

Zappavigna, M. (2011). Ambient affiliation: A linguistic perspective on Twitter. NewMedia and Society, 13, 788–806.

Zhang, W., Xu, H., & Wan, W. (2012). Weakness finder: Find product weakness fromChinese reviews by using aspects based sentiment analysis. Expert Systems withApplications, 39, 10283–10291.

Zhang, C., Zeng, D., Li, J., Wang, F., & Zuo, W. (2009). Sentiment analysis of Chinesedocuments: From sentence to document level. Journal of the American Society forInformation Science and Technology, 60, 2474–2487.

Zhuang, L., Jing, F., & Zhu, X. (2006). Movie review mining and summarization.Proceedings of the 15th ACM conference on information and knowledgemanagement (pp. 43–50). New York, NY.

expert systems with applications - ایران...

Documents